A deep learning-enriched framework for analyzing brain functional connectivity.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Brain function relies on directed communication between regions, utilizing asymmetric connections.
• Deep learning models excel at decoding brain states but struggle with characterizing information flow in functional networks.
• Analyzing spectral directed functional connectivity is crucial for understanding complex brain processes.

Purpose of the Study:

• To develop a deep learning-enriched framework for analyzing spectral directed functional connectivity.
• To create novel, non-linear inflow and outflow measures for brain networks.
• To identify informative frequency content and connectivity patterns for brain states.

Main Methods:

• Designed 'Functional-Connectivity-Net' (FCNet), an interpretable convolutional neural network, trained to discriminate brain states from functional connectivity.
• Utilized DeepLIFT for explaining network decisions and identifying key frequency and connectivity features.
• Applied the framework to EEG functional connectivity data from scalp and cortex during motor imagery tasks.

Main Results:

• FCNet's explanations align with known spectral connectivity changes during motor imagery.
• Novel network-based measures effectively capture connectivity changes, comparable to graph theory metrics.
• The framework successfully identified informative frequencies and connectivity inflow/outflows.

Conclusions:

• The proposed deep learning framework enhances the analysis of spectral directed functional connectivity.
• It provides valuable insights into the predictability and informative components of brain functional networks.
• This approach aids in understanding brain states and information flow during cognitive and motor tasks.